Quick intro to calculus with Julia

Here are some Julia usages to create calculus objects.

The Julia packages loaded below are all loaded when the CalculusWithJulia package is loaded.

A Julia package is loaded with the using command:

using LinearAlgebra

The LinearAlgebra package comes with a Julia installation. Other packages can be added. Something like:

using Pkg
Pkg.add("SomePackageName")

These notes have an accompanying package, CalculusWithJulia, that when installed, as above, also installs most of the necessary packages to perform the examples.

Packages need only be installed once, but they must be loaded into each session for which they will be used.

using CalculusWithJulia

Packages can also be loaded through import PackageName. Importing does not add the exported objects of a function into the namespace, so is used when there are possible name collisions.

Types

Objects in Julia are "typed." Common numeric types are Float64, Int64 for floating point numbers and integers. Less used here are types like Rational{Int64}, specifying rational numbers with a numerator and denominator as Int64; or Complex{Float64}, specifying a comlex number with floating point components. Julia also has BigFloat and BigInt for arbitrary precision types. Typically, operations use "promotion" to ensure the combination of types is appropriate. Other useful types are Function, an abstract type describing functions; Bool for true and false values; and Sym for symbolic values (through SymPy).

For the most part the type will not be so important, but it is useful to know that for some function calls the type of the argument will decide what method ultimately gets called. (This allows symbolic types to interact with Julia functions in an idiomatic manner.)

Functions

Definition

Functions can be defined four ways:

one statement functions follow traditional mathematics notation:

f(x) = exp(x) * 2x

f (generic function with 1 method)

multi-statement functions are defined with the function keyword. The end statement ends the definition. The last evaluated command is returned. There is no need for explicit return statement, though it can be useful for control flow.

function g(x)
  a = sin(x)^2
  a + a^2 + a^3
end

g (generic function with 1 method)

Anonymous functions are useful for example, as arguments to other functions or as return values.

using ForwardDiff   # add-on package to find numeric derivatives using automatic differentiation
D(f::Function) = x -> ForwardDiff.derivative(f, x)

D (generic function with 1 method)

The above turns a function that finds the derivative of f at a point and creates a function, D, returning a derivative function.

Anonymous function may also be created using the function keyword.

For mathematical functions $f: R^n \rightarrow R^m$ when $n$ or $m$ is bigger than 1 we have:

\[ n =1 \]

by $m > 1$ use a "vector" for the return value

r(t) = [sin(t), cos(t), t]

r (generic function with 1 method)

\[ n > 1 \]

by $m=1$ use multiple arguments

f(x,y,z) = x*y + y*z + z*x
f(v) = f(v...)

f (generic function with 2 methods)

Some functions need to pass in a container of values, for this the last definition is useful to expand the values. Splatting takes a container and treats the values like individual arguments.

Alternatively, indexing can be used directly, as in:

f(x) = x[1]*x[2] + x[2]*x[3] + x[3]*x[1]

f (generic function with 2 methods)

For vector fields ($n,m > 1$) a combination is used:

F(x,y,z) = [-y, x, z]
F(v) = F(v...)

F (generic function with 2 methods)

Calling a function

Functions are called using parentheses to group the arguments.

f(t) = sin(t)*sqrt(t)
sin(1), sqrt(1), f(1)

(0.8414709848078965, 1.0, 0.8414709848078965)

Multiple dispatch

Julia can have many methods for a single generic function. (E.g., it can have many different implementations of addiion when the + sign is encountered.) The type of an argument and the number of arguments are used for dispatch.

Here the number of arguments is used:

Area(w, h) = w * h    # area of rectangle
Area(w) = Area(w, w)  # area of square using area of rectangle defintion

Area (generic function with 2 methods)

Calling Area(5) will call Area(5,5) which will return 5*5.

Similarly, the definition for a vector field:

F(x,y,z) = [-y, x, z]
F(v) = F(v...)

F (generic function with 2 methods)

takes advantage of multiple dispatch to allow either a vector argument or individual arguments.

Type parameters can be used to restrict the type of arguments that are permitted. The D(f::Function) definition illustrates how the D function is restricted to Function objects.

Keyword arguments

Optional arguments may be specified with keywords, when the function is defined to use them. Keywords are separated from positional arguments using a semicolon, ;:

circle(x; r=1) = sqrt(r^2 - x^2)
circle(0.5), circle(0.5, r=10)

(0.8660254037844386, 9.987492177719089)

Symbolic objects

The add-on SymPy package allows for symbolic expressions to be used. Symbolic values are defined with @vars, as below.

using SymPy
@vars x y z  # no comma
x^2 + y^3 + z

\begin{equation*}x^{2} + y^{3} + z\end{equation*}

Assumptions on the variables can be useful, particularly with simplification, as in

@vars x y z real=true

(x, y, z)

Symbolic expressions flow through Julia functions symbolically

sin(x)^2 + cos(x)^2

\begin{equation*}\sin^{2}{\left (x \right )} + \cos^{2}{\left (x \right )}\end{equation*}

Numbers are symbolic once SymPy interacts with them:

x - x + 1  # 1 is now symbolic

\begin{equation*}1\end{equation*}

The number PI is a symbolic pi. a

sin(PI), sin(pi)

(0, 1.2246467991473532e-16)

Use Sym to create symbolic numbers, N to find a Julia number from a symbolic number:

1 / Sym(2)

\begin{equation*}\frac{1}{2}\end{equation*}

N(PI)

π = 3.1415926535897...

Many Julia functions will work with symbolic objects through multiple dispatch (e.g., sin, cos, ...). Sympy functions that are not in Julia can be accessed through the sympy object using dot-call notation:

sympy.harmonic(10)

\begin{equation*}\frac{7381}{2520}\end{equation*}

Containers

We use a few different containers:

Tuples. These are objects grouped together using parentheses. They need not be of the same type

x1 = (1, "two", 3.0)

(1, "two", 3.0)

Tuples are useful for programming. For example, they are uesd to return multiple values from a function.

Vectors. These are objects of the same type (typically) grouped together using square brackets, values separated by commas:

x2 = [1, 2, 3.0]  # 3.0 makes theses all floating point

3-element Array{Float64,1}:
 1.0
 2.0
 3.0

Matrices. Like vectors, these have the same type, only they are 2-dimensional. Use spaces to separate values along a row; semicolons to separate rows:

x3 = [1 2 3; 4 5 6; 7 8 9]

3×3 Array{Int64,2}:
 1  2  3
 4  5  6
 7  8  9

Row vectors. A vector is 1 dimensional, though it may be identified as a column of two dimensional matrix. A row vector is a two-dimensional matrix with a single row:

x4 = [1 2 3.0]

1×3 Array{Float64,2}:
 1.0  2.0  3.0

These have indexing using square brackets:

x1[1], x2[2], x3[3]

(1, 2.0, 7)

Matrices are usually indexed by row and column:

x3[1,2] # row one column two

For vectors and matrices – but not tuples – indexing can be used to change a value in the container:

x2[1], x3[1,1] = 2, 2

(2, 2)

Vectors and matrices are arrays. Arrays have mathematical operations, such as addition and subtraction, defined for them. Tuples do not.

Destructuring is another way to get at the entries:

a,b,c = x2

3-element Array{Float64,1}:
 2.0
 2.0
 3.0

structured collections

An arithmetic progression, $a, a+h, a+2h, ..., b$ can be produced efficiently using the range operator:

5:10:55  # an object that describes 5, 15, 25, 35, 45, 55

5:10:55

If h=1 it can be omitted:

1:10     # an object that describes 1,2,3,4,5,6,7,8,9,10

1:10

The range function can efficiently describe $n$ evenly spaced points between a and b:

range(0, pi, length=5)  # range(a, stop=b, length=n) for version 1.0

0.0:0.7853981633974483:3.141592653589793

This is useful for creating regularly spaced values needed for certain plots.

Iteration

The for keyword is useful for iteration, Here is a traditional for loop, as i loops over each entry of the vector [1,2,3]:

for i in [1,2,3]
  print(i)
end

List comprehensions are similar, but are useful as they perform the iteration and collect the values:

[i^2 for i in [1,2,3]]

3-element Array{Int64,1}:
 1
 4
 9

Comprehesions can also be used to make matrices

[1/(i+j) for i in 1:3, j in 1:4]

3×4 Array{Float64,2}:
 0.5       0.333333  0.25      0.2     
 0.333333  0.25      0.2       0.166667
 0.25      0.2       0.166667  0.142857

Comprehensions apply an expression to each entry in a container through iteration. Applying a function to each entry of a container can be facilitated by:

Broadcasting. Using . before an operation instructs Julia to match up sizes (possibly extending to do so) and then apply the operation element by element:

xs = [1,2,3]
sin.(xs)   # sin(1), sin(2), sin(3)
bases = [5,5,10]
log.(bases, xs)  # log(5, 1), log(5,2), log(10, 3)

3-element Array{Float64,1}:
 0.0                
 0.43067655807339306
 0.47712125471966244

Row and column vectors can fill in:

ys = [4 5] # a row vector
f(x,y) = (x,y)
f.(xs, ys)    # broadcasting a column and row vector makes a matrix, then applies f.

3×2 Array{Tuple{Int64,Int64},2}:
 (1, 4)  (1, 5)
 (2, 4)  (2, 5)
 (3, 4)  (3, 5)

The map function applies a function to each element:

map(sin, [1,2,3])

3-element Array{Float64,1}:
 0.8414709848078965
 0.9092974268256817
 0.1411200080598672

Plots

The following commands use the Plots package. The Plots package expects a choice of backend. We will use both plotly and gr (and occasionally pyplot()).

using Plots
plotly()      # select plotly. Use `gr()` for GR

Plots.PlotlyBackend()

Plotting a univariate function $f:R \rightarrow R$

using plot(f, a, b)

plot(sin, 0, 2pi)

f(x) = exp(-x/2pi)*sin(x)
plot(f, 0, 2pi)

Or with an anonymous function

plot(x -> sin(x) + sin(2x), 0, 2pi)

note("""The time to first plot can be lengthy!. This can be removed by creating a custom `Julia` image, but that is no introductory level stuff. As well, standalone plotting packages offer quicker first plots, but the simplicity of `Plots` is preferred. Subsequent plots are not so time consuming, as the initial time is spent compiling functions so their re-use is speedy.
""")

The time to first plot can be lengthy!. This can be removed by creating a custom Julia image, but that is no introductory level stuff. As well, standalone plotting packages offer quicker first plots, but the simplicity of Plots is preferred. Subsequent plots are not so time consuming, as the initial time is spent compiling functions so their re-use is speedy.

Arguments of interest include

Attribute	Value
`legend`	A boolean, specify `false` to inhibit drawing a legend
`aspect_ratio`	Use `:equal` to have x and y axis have same scale
`linewidth`	Ingters greater than 1 will thicken lines drawn
`color`	A color may be specified by a symbol (leading `:`).
	E.g., `:black`, `:red`, `:blue`

using plot(xs, ys)

The lower level interface to plot involves creating x and y values to plot:

xs = range(0, 2pi, length=100)
ys = sin.(xs)
plot(xs, ys, color=:red)

plotting a symbolic expression

A symbolic expression of single variable can be plotted as a function is:

@vars x
plot(exp(-x/2pi)*sin(x), 0, 2pi)

Multiple functions

The ! Julia convention to modify an object is used by the plot command, so plot! will add to the existing plot:

plot(sin, 0, 2pi, color=:red)
plot!(cos, 0, 2pi, color=:blue)
plot!(zero, color=:green)  # no a, b then inherited from graph.

The zero function is just 0 (useful when the type of a number is important).

Plotting a parameterized (space) curve function $f:R \rightarrow R^n$, $n = 2$ or $3$

Using plot(xs, ys)

Let $f(t) = e^{t/2\pi} \langle \cos(t), \sin(t)\rangle$ be a parameterized function. Then the $t$ values can be generated as follows:

ts = range(0, 2pi, length = 100)
xs = [exp(t/2pi) * cos(t) for t in ts]
ys = [exp(t/2pi) * sin(t) for t in ts]
plot(xs, ys)

using plot(f1, f2, a, b). If the two functions describing the components are available, then

f1(t) = exp(t/2pi) * cos(t)
f2(t) = exp(t/2pi) * sin(t)
plot(f1, f2, 0, 2pi)

Using plot_parametric_curve. If the curve is described as a function of t with a vector output, then the CalculusWithJulia package provides plot_parametric_curve to produce a plot:

r(t) = exp(t/2pi) * [cos(t), sin(t)]
plot_parametric_curve(r, 0, 2pi)

The low-level approach doesn't quite work as easily as desired:

ts = range(0, 2pi, length = 4)
vs = r.(ts)

4-element Array{Array{Float64,1},1}:
 [1.0, 0.0]             
 [-0.697806, 1.20864]   
 [-0.973867, -1.68679]  
 [2.71828, -6.65787e-16]

As seen, the values are a vector of vectors. To plot a reshaping needs to be done:

ts = range(0, 2pi, length = 100)
vs = r.(ts)
xs = [vs[i][1] for i in eachindex(vs)]
ys = [vs[i][2] for i in eachindex(vs)]
plot(xs, ys)

This approach is faciliated by the unzip function in CalculusWithJulia:

plot(unzip(vs)...)

Plotting an arrow

An arrow in 2D can be plotted with the quiver command. We show the arrow(p, v) (or arrow!(p,v) function) from the CalculusWithJulia package, which has an easier syntax:

gr()
plot_parametric_curve(r, 0, 2pi)
t0 = pi/8
arrow!(r(t0), r'(t0))

The GR package makes nicer arrows that Plotly.

Plotting a scalar function $f:R^2 \rightarrow R$

The surface and contour functions are available to visualize a scalar function of $2$ variables:

A surface plot

(The plotly backend allows for rotation by the mouse.)

plotly()
f(x, y) = 2 - x^2 + y^2
xs = ys = range(-2,2, length=25)
surface(xs, ys, f)

The function generates the $z$ values, this can be done by the user and then passed to the surface(xs, ys, zs) format:

surface(xs, ys, f.(xs', ys))

A contour plot

The contour function is like the surface function.

contour(xs, ys, f)

contour(xs, ys, f.(xs', ys))

An implicit equation. The constraint $f(x,y)=c$ generates an implicit equation. While contour can be used for this plot by adjusting the requested contours, the ImplicitEquations package can as well, and perhaps is easier. This package is loaded with CalculusWithJulia; loading it by itself will lead to naming conflicts with SymPy, so best not to do so. ImplicitEquations plots predicates formed by Eq, Le, Lt, Ge, and Gt (or some unicode counterparts). For example to plot when $f(x,y) = \sin(xy) - \cos(xy) \leq 0$ we have:

f(x,y) = sin(x*y) - cos(x*y)
plot(Le(f, 0))

Plotting a parameterized surface $f:R^2 \rightarrow R^3$

The plotly (and pyplot) backends allow plotting of parameterized surfaces. The plot_parametric_surface function from CalculusWithJulia handles the details:

plotly()
Phi(theta, phi) = [sin(phi) * cos(theta), sin(phi) * sin(theta), cos(phi)]
plot_parametric_surface(Phi, xlim=(0,pi/4), ylim=(0,pi))

The low-level surface(xs,ys,zs) is used, and can be specified directly as follows:

X(theta, phi) = sin(phi)*cos(theta)
Y(theta, phi) = sin(phi)*sin(theta)
Z(theta, phi) = cos(phi)
thetas = range(0, pi/4, length=20)
phis = range(0, pi, length=20)
surface(X.(thetas, phis'), Y.(thetas, phis'), Z.(thetas, phis'))

Plotting a vector field $F:R^2 \rightarrow R^2$. The CalculusWithJulia package provides vectorfieldplot, used as:

gr()  # better arrows
F(x,y) = [-y, x]
vectorfieldplot(F, xlim=(-2, 2), ylim=(-2,2), nx=10, ny=10)

There is also vectorfieldplot3d.

Limits

Limits can be investigated numerically by forming tables, eg.:

xs = [1, 1/10, 1/100, 1/1000]
f(x) = sin(x)/x
[xs f.(xs)]

4×2 Array{Float64,2}:
 1.0    0.841471
 0.1    0.998334
 0.01   0.999983
 0.001  1.0

Symbolically, SymPy provides a limit function:

@vars x
limit(sin(x)/x, x => 0)

\begin{equation*}1\end{equation*}

@vars h x
limit((sin(x+h) - sin(x))/h, h => 0)

\begin{equation*}\cos{\left (x \right )}\end{equation*}

Derivatives

There are numeric and symbolic approaches to derivatives. For the numeric approach we use the ForwardDiff package, which performs automatic differentiation.

Derivatives of univariate functions

Numerically, the ForwardDiff.derivative(f, x) function call will find the derivative of the function f at the point x:

ForwardDiff.derivative(sin, pi/3) - cos(pi/3)

0.0

The CalculusWithJulia package overides the ' (adjoint) syntax for functions to provide a derivative which takes a function and returns a function, so its usage is familiar

f(x) = sin(x)
f'(pi/3) - cos(pi/3)  # or just sin'(pi/3) - cos(pi/3)

0.0

Higher order derivatives are possible as well,

f(x) = sin(x)
f''''(pi/3) - f(pi/3)

0.0

Symbolically, the diff function of SymPy finds derivatives.

@vars x
f(x) = exp(-x)*sin(x)
ex = f(x)  # symbolic expression
diff(ex, x)   # or just diff(f(x), x)

\begin{equation*}- e^{- x} \sin{\left (x \right )} + e^{- x} \cos{\left (x \right )}\end{equation*}

Higher order derivatives can be specified as well

diff(ex, x, x)

\begin{equation*}- 2 e^{- x} \cos{\left (x \right )}\end{equation*}

Or with a number:

diff(ex, x, 5)

\begin{equation*}4 \left(\sin{\left (x \right )} - \cos{\left (x \right )}\right) e^{- x}\end{equation*}

The variable is important, as this allows parameters to be symbolic

@vars mu sigma x
diff(exp(-((x-mu)/sigma)^2/2), x)

\begin{equation*}- \frac{\left(- 2 \mu + 2 x\right) e^{- \frac{\left(- \mu + x\right)^{2}}{2 \sigma^{2}}}}{2 \sigma^{2}}\end{equation*}

partial derivatives

There is no direct partial derivative function provided by ForwardDiff, rather we use the result of the ForwardDiff.gradient function, which finds the partial derivatives for each variable. To use this, the function must be defined in terms of a point or vector.

f(x,y,z) = x*y + y*z + z*x
f(v) = f(v...)               # this is needed for ForwardDiff.gradient
ForwardDiff.gradient(f, [1,2,3])

3-element Array{Int64,1}:
 5
 4
 3

We can see directly that $\partial{f}/\partial{x} = \langle y + z\rangle$. At the point $(1,2,3)$, this is $5$, as returned above.

Symbolically, diff is used for partial derivatives:

@vars x y z
ex = x*y + y*z + z*x
diff(ex, x)    # ∂f/∂x

\begin{equation*}y + z\end{equation*}

Gradient

As seen, the ForwardDiff.gradient function finds the gradient at a point. In CalculusWithJulia, the gradient is extended to returns a function when called with no additional arguments:

f(x,y,z) =  x*y + y*z + z*x
f(v) = f(v...)
gradient(f)(1,2,3) - gradient(f, [1,2,3])

3-element Array{Int64,1}:
 0
 0
 0

The ∇ symbol, formed by entering \nabla[tab], is mathematical syntax for the gradient, and is defined in CalculusWithJulia.

∇(f)(1,2,3)   # same as gradient(f, [1,2,3])

3-element Array{Int64,1}:
 5
 4
 3

In SymPy, there is no gradient function, though finding the gradient is easy through broadcasting:

@vars x y z
ex = x*y + y*z + z*x
diff.(ex, [x,y,z])  # [diff(ex, x), diff(ex, y), diff(ex, z)]

\[ \left[ \begin{array}{r}y + z\\x + z\\x + y\end{array} \right] \]

The CalculusWithJulia package provides a method for gradient:

gradient(ex, [x,y,z])

\[ \left[ \begin{array}{r}y + z\\x + z\\x + y\end{array} \right] \]

The ∇ symbol is an alias. It can guess the order of the free symbols, but generally specifying them is needed. This is done with a tuple:

∇((ex, [x,y,z]))

\[ \left[ \begin{array}{r}y + z\\x + z\\x + y\end{array} \right] \]

Jacobian

The Jacobian of a function $f:R^n \rightarrow R^m$ is a $m\times n$ matrix of partial derivatives. Numerically, ForwardDiff.jacobian can find the Jacobian of a function at a point:

F(u,v) = [u*cos(v), u*sin(v), u]
F(v) = F(v...)    # needed for ForwardDiff.jacobian
pt = [1, pi/4]
ForwardDiff.jacobian(F , pt)

3×2 Array{Float64,2}:
 0.707107  -0.707107
 0.707107   0.707107
 1.0        0.0

Symbolically, the jacobian function is a method of a matrix, so the calling pattern is different. (Of the form object.method(arguments...).)

@vars u v
ex = F(u,v)
ex.jacobian([u,v])

\[\left[ \begin{array}{rr}\cos{\left (v \right )}&- u \sin{\left (v \right )}\\\sin{\left (v \right )}&u \cos{\left (v \right )}\\1&0\end{array}\right]\]

As the Jacobian can be identified as the matrix with rows given by the transpose of the gradient of the component, it can be computed directly, but it is more difficult:

@vars u v real=true
vcat([diff.(ex, [u,v])' for ex in F(u,v)]...)

3×2 Array{Any,2}:
 cos(v)  -u*sin(v)
 sin(v)   u*cos(v)
      1          0

Divergence

Numerically, the divergence can be computed from the Jacobian by adding the diagonal elements. This is a numerically inefficient, as the other partial derivates must be found and discarded, but this is generally not an issue for these notes. The following uses tr (the trace from the LinearAlgebra package) to find the sum of a diagonal.

F(x,y,z) = [-y, x, z]
F(v) = F(v...)
pt = [1,2,3]
tr(ForwardDiff.jacobian(F , pt))

The CalculusWithJulia package provides divergence to compute the divergence and provides the ∇ ⋅ notation (\nabla[tab]\cdot[tab]):

divergence(F, [1,2,3])
(∇⋅F)(1,2,3)    # not ∇⋅F(1,2,3) as that evaluates F(1,2,3) before the divergence

1.0

Symbolically, the divergence can be found directly:

@vars x y z
ex = F(x,y,z)
sum(diff.(ex, [x,y,z]))    # sum of [diff(ex[1], x), diff(ex[2],y), diff(ex[3], z)]

\begin{equation*}1\end{equation*}

The divergence function can be used for symbolic expressions:

divergence(ex, [x,y,z])
∇⋅(F(x,y,z), [x,y,z])

\begin{equation*}1\end{equation*}

Curl

The curl can be computed from the off-diagonal elements of the Jacobian. The calculation follows the formula. The CalculusWithJulia package provides curl to compute this:

F(x,y,z) = [-y, x, 1]
F(v) = F(v...)
curl(F, [1,2,3])

3-element Array{Float64,1}:
  0.0
 -0.0
  2.0

As well, if no point is specified, a function is returned for which a point may be specified using 3 coordinates or a vector

curl(F)(1,2,3), curl(F)([1,2,3])

([0.0, -0.0, 2.0], [0.0, -0.0, 2.0])

Finally, the ∇ × (\nabla[tab]\times[tab] notation is available)

(∇×F)(1,2,3)

3-element Array{Float64,1}:
  0.0
 -0.0
  2.0

Integrals

Numeric integration is provided by the QuadGK package, for univariate integrals, and the HCubature package for higher dimensional integrals.

Integrals of univariate functions

A definite integral may be computed numerically using quadgk

using QuadGK
quadgk(sin, 0, pi)

(2.0, 1.7905676941154525e-12)

The answer and an estimate for the worst case error is returned.

If singularities are avoided, improper integrals are computed as well:

quadgk(x->1/x^(1/2), 0, 1)

(1.9999999845983916, 2.3762511924588765e-8)

SymPy provides the integrate function to compute both definite and indefinite integrals.

@vars a x real=true
integrate(exp(a*x)*sin(x), x)

\begin{equation*}\frac{a e^{a x} \sin{\left (x \right )}}{a^{2} + 1} - \frac{e^{a x} \cos{\left (x \right )}}{a^{2} + 1}\end{equation*}

Like diff the variable to integrate is specified.

Definite integrals use a tuple, (variable, a, b) to integrate:

integrate(sin(a + x), (x, 0, PI))   # \int_0^PI sin(a+x) dx

\begin{equation*}2 \cos{\left (a \right )}\end{equation*}

2D and 3D iterated integrals

Two and three dimensional integrals over box-like regions are computed numerically with the hcubature function from the HCubature package. If the box is $[x_1, y_1]\times[x_2,y_2]\times\cdots\times[x_n,y_n]$ then the limits are specified through tuples of the form $(x_1,x_2,\dots,x_n)$ and $(y_1,y_2,\dots,y_n)$.

f(x,y) = x*y^2
f(v) = f(v...)
hcubature(f, (0,0), (1, 2))  # computes \int_0^1\int0^2 f(x,y) dy dx

(1.333333333333333, 4.440892098500626e-16)

The calling pattern for more dimensions is identical.

f(x,y,z) = x*y^2*z^3
f(v) = f(v...)
hcubature(f, (0,0,0), (1, 2,3))

(27.0, 0.0)

The box-like region requirement means a change of variables may be necessary. For example, to integrate over the region $x^2 + y^2 \leq 1; x \geq 0$, polar coordinates can be used with $(r,\theta)$ in $[0,1]\times[-\pi/2,\pi/2]$. When changing variables, the Jacobian enters into the formula, through

\[ ~ \iint_{G(S)} f(\vec{x}) dV = \iint_S (f \circ G)(\vec{u}) |\det(J_G)(\vec{u})| dU. ~ \]

Here we implement this:

f(x,y) = x*y^2
f(v) = f(v...)
Phi(r, theta) = r * [cos(theta), sin(theta)]
Phi(rtheta) = Phi(rtheta...)
integrand(rtheta) = f(Phi(rtheta)) * det(ForwardDiff.jacobian(Phi, rtheta))
hcubature(integrand, (0.0,-pi/2), (1.0, pi/2))

(0.13333333333904923, 1.9853799966359355e-9)

In CalculusWithJulia a fubini function is provided to compute numeric integrals over regions which can be described by curves represented by functions. E.g., for this problem:

fubini(f, (x -> -sqrt(1-x^2), x -> sqrt(1-x^2)), (0, 1))

0.13333333332775762

This function is for convenience, but is not performant.

Symbolically, the integrate function allows additional terms to be specified. For example, the above could be done through:

@vars x y real=true
integrate(x * y^2, (y, -sqrt(1-x^2), sqrt(1-x^2)), (x, 0, 1))

\begin{equation*}\frac{2}{15}\end{equation*}

Line integrals

A line integral of $f$ parameterized by $\vec{r}(t)$ is computed by:

\[ ~ \int_a^b (f\circ\vec{r})(t) \| \frac{dr}{dt}\| dt. ~ \]

For example, if $f(x,y) = 2 - x^2 - y^2$ and $r(t) = 1/t \langle \cos(t), \sin(t) \rangle$, then the line integral over $[1,2]$ is given by:

f(x,y) = 2 - x^2 - y^2
f(v) = f(v...)
r(t) = [cos(t), sin(t)]/t
integrand(t) = (f∘r)(t) * norm(r'(t))
quadgk(integrand, 1, 2)

(1.2399213772953277, 4.525271268818187e-9)

To integrate a line integral through a vector field, say $\int_C F \cdot\hat{T} ds=\int_C F\cdot \vec{r}'(t) dt$ we have, for example,

F(x,y) = [-y, x]
F(v) = F(v...)
r(t) =  [cos(t), sin(t)]/t
integrand(t) = (F∘r)(t) ⋅ r'(t)
quadgk(integrand, 1, 2)

(0.5, 2.1134927141730486e-10)

Symbolically, there is no real difference from a 1-dimensional integral. Let $\phi = 1/\|r\|$ and integrate the gradient field over one turn of the helix $\vec{r}(t) = \langle \cos(t), \sin(t), t\rangle$.

@vars x y z t real=true
phi(x,y,z) = 1/sqrt(x^2 + y^2 + z^2)
r(t) = [cos(t), sin(t), t]
∇phi = diff.(phi(x,y,z), [x,y,z])
∇phi_r = subs.(∇phi, x.=> r(t)[1], y.=>r(t)[2], z.=>r(t)[3])
rp = diff.(r(t), t)
ex = simplify(∇phi_r ⋅ rp )

\begin{equation*}- \frac{t}{\left(t^{2} + 1\right)^{\frac{3}{2}}}\end{equation*}

Then

integrate(ex, (t, 0, 2PI))

\begin{equation*}-1 + \frac{1}{\sqrt{1 + 4 \pi^{2}}}\end{equation*}

surface integrals

The surface integral for a parameterized surface involves a surface element $\|\partial\Phi/\partial{u} \times \partial\Phi/\partial{v}\|$. This can be computed numerically with:

Phi(u,v) = [u*cos(v), u*sin(v), u]
Phi(v) = Phi(v...)
function SE(Phi, pt)
   J = ForwardDiff.jacobian(Phi, pt)
   J[:,1] × J[:,2]
end
norm(SE(Phi, [1,2]))

1.4142135623730951

To find the surface integral ($f=1$) for this surface over $[0,1] \times [0,2pi]$, we have:

hcubature(pt -> norm(SE(Phi, pt)), (0.0,0.0), (1.0, 2pi))

(4.442882938158366, 2.6645352591003757e-15)

Symbolically, the approach is similar:

@vars u v real=true
ex = Phi(u,v)
J = ex.jacobian([u,v])
SurfEl = norm(J[:,1] × J[:,2]) |> simplify

\begin{equation*}\sqrt{2} \left|{u}\right|\end{equation*}

Then

integrate(SurfEl, (u, 0, 1), (v, 0, 2PI))

\begin{equation*}\sqrt{2} \pi\end{equation*}